Missing GC root in DecisionTrees

[Keno]

Reported in https://discourse.julialang.org/t/fatal-error-in-type-inference-type-bound-in-0-7-rc2

It looks like what's happening is that the root for numPerScore in the bagged_tree function is getting dropped and that array is then freed during the createTrnValStatsForThisIteration function, but used later.

[Keno]

What looks to be happening here is that the recently enabled SLP vectorizer is vectorizing the load of a tracked pointer, causing our GC root lowering to lose track of it.

[Keno]

Minimal reproducer:

mutable struct foo
    x::Int
end

@noinline make_foo() = (foo(1), foo(rand(1:10)), foo(rand(1:10)))
@noinline function use_tuple(c)
    @Base.show c
end

function repackage()
    (_, a, b) = make_foo()
    GC.gc()
    c = (foo(1), foo(2), a, b)
    use_tuple(c)
    nothing
end

julia> repackage()
c = (foo(1), foo(2), Tuple{Core.Compiler.UseRef,Nothing}(Core.Compiler.UseRef(:((Base.getfield)(_3, 2)), 4), nothing), Tuple{Core.Compiler.UseRef,Nothing}(Core.Compiler.UseRef(:(%new(Base.ImmutableDict{Symbol,Any}, %2, %3, %4)), 5), nothing))